Adversarial training for low-complexity convolutional neural networks using in spectrum sensing

Spectrum sensing of orthogonal frequency division multiplex (OFDM) system has always been a challenge in cognitive radios (CR). In this paper on the basis of “classification converted sensing” scheme, the cyclostationary periodogram generated by OFDM pilots is deduced in the form of images. These images are then plugged into the convolutional neural networks (CNNs) for classifications due to CNN's strength in image classification. More importantly, certain of concerns about CNN adoption in CR system is settled. Firstly, to achieve spectrum sensing against severe noise pollution and channel fading, we use the adversarial training where a CR-specific, modified training database is proposed. Then, to settle the serviceability which is constrained by the computing power at the CR user end, the input images and the CNN architecture are refined to guarantee a low-complexity but high-performance sensing scheme. Simulation results proved our method possesses an excellent sensing capability while achieving higher detection accuracy over the conventional way.